NVIDIA NVENC NVDEC Matrix - Elpamsoft.com

1 Linux Plex Media Server Cheat Sheet (26-April-2019) No warranty. Use at own risk. Results may vary. Some information from 3rd party sources. No copyrights implied. See your doctor if problems persist. GPU Transcoding using NVIDIA NVENC and NVENC Graphics Card Model Chipset Family Chip Model # of Chips Session Limit NVENC FPS NVDEC FPS Streams for VRAM* VRAM Streams Card Load * (Less than 100% Best) Recommended Plex Streams VDPAU Notes Chips h264 720p h264 1080p HEVC Chips H264 HEVC VRAM GeForce GT 630 - 640 GeForce GT 710 - 730 Kepler GK208 1 2 1 157-490 78-220 - 1 161 - 1GB 4 74% 4 D NVENC : No H264 4:4:4 or Lossless on Kepler K620 Can be GK107 or GM107 GeForce GT 630 - 640 GeForce GTX 650, GT 740 GK107 1 2 1 157-490 78-220 - 1 161 - 1-2GB 4-8 74% - 149% 4-5 quadro K420, K600 1 2 1 157-490 78-220 - 1 161 - 1GB 4 74% 4 quadro K620 1 2 1 157-490 78-220 - 1 161 - 1-2GB 4-8 74% - 149% 4-5 quadro K2000 1 U 1 157-490 78-220 - 1 161 - 2GB 8 149% 5 GeForce GTX 645 -660 Ti Boost GeForce GT 740 GK106 1 2 1 157-490 78-220 - 1 161 - 1-2GB 4-8 74% - 149% 4-5 quadro K4000 1 U 1 157-490 78-220 - 1 161 - 3GB 12 224% 5 GeForce GTX 660 - 690 GeForce GTX 760 - 770 GK104 1 2 1 157-490 78-220 - 1 161 - 2-8GB 8-32 149% - 596% 5 quadro K4200-K5000 1 U 1 157-490 78-220 - 1 161 - 4-8GB 16-32 298% - 596% 5 GeForce GT 780 - 780 Ti Kepler (2nd Gen)

2 GK110 1 2 1 157-490 78-220 - 1 161 - 3-6GB 12-25 224% - 466% 5 Kepler 2nd Gen GeForce GTX Titan / Titan Black 1 2 1 157-490 78-220 - 1 161 - 6GB 25 466% 5 GeForce GTX Titan Z 2 2 2 157-490 156-440 - 2 322 - 6GB 25 233% 10 Double NVENC / NVDEC Chips quadro K6000 GK110B 1 U 1 157-490 78-220 - 1 161 - 12GB 49 913% 5 GeForce GTX 745 - 750 Ti Maxwell (1st Gen) GM107 1 2 1 502-826 211-345 - 1 376 - 2-4GB 8-16 70% 8-12 E PureVideo hardware and software running on the shader array to decode HEVC ( ) as partial/hybrid hardware video decoding. quadro K1200 1 2 1 502-826 211-345 - 1 376 - 4GB 16 139% 12 quadro K2200 1 U 1 502-826 211-345 - 1 376 - 4GB 16 139% 12 Telsa M10 1 U 1 502-826 211-345 - 1 376 - 8GB 32 287% 12 GeFroce 830M-940MX, MX110-130 GM108 1 0 0 - - - 0 - - 1-2GB 0 - - No NVENC / NVDEC Chips quadro K620M 1 0 0 - - - 0 - - 2GB 0 - - GeForce GTX 750, 950 - 960 Maxwell (2nd Gen) GM206 1 2 1 641-1111 261-432 142-200 1 376 408 1-3GB 4-12 32% - 96% 4-12 F Dedicated HEVC Main (8-bit) & Main 10 (10-bit) and VP9 hardware decoding video decoding up to 4K quadro M2000 1 U 1 641-1111 261-432 142-200 1 376 408 4GB 16 128% 12 F GeForce GTX 960 Ti - 980 GM204 1 2 2 522-864 284-400 1 376 ?

3 2-8GB 8-33 64%-263% 8-12 E PureVideo hardware and software running on the shader array to decode HEVC ( ) as partial/hybrid hardware video decoding. Expect Slow HEVC using CPU quadro M4000-M5000 Series 1 U 2 522-864 284-400 1 376 ? 8GB 33 263% 12 GeForce GTX 980 Ti GM200 1 2 2 522-864 284-400 1 376 ? 6GB 25 199% 12 GeForce GTX Titan X 1 2 2 522-864 284-400 1 376 ? 12GB 50 391% 12 quadro M6000 Series 1 U 2 522-864 284-400 1 376 ? 12GB 50 391% 12 GeForce GT 1030 Pascal GP108 1 0 0 - - 1 658 720 2GB 0 - - H No NVENC Support GeForce GTX 1050 / 1050 Ti Pascal GP107 1 2 1 388-631 259-395 1 658 720 4GB 14 67% 14 Pascal adds 8K NVDEC Support and 8K NVENC on some cards?? quadro P400-P620 1 2 1 388-631 259-395 1 658 720 2GB 7 33% 7 quadro P1000 1 2 1 388-631 259-395 1 658 720 4GB 14 67% 14 GeForce GTX 1050 / 1050 Ti GP106 1 2 1 388-631 259-395 1 658 720 2-4GB 7-14 33% - 67% 7-14 GeForce GTX 1060 1 2 1 388-631 259-395 1 658 720 3-6GB 10-20 48% - 95% 10-20 quadro P2000 1 U 1 388-631 259-395 1 658 720 5GB 17 81% 17 GeForce GTX 1060 GP104 1 2 1 388-631 259-395 1 658 720 3-6GB 10-20 48% - 95% 10-20 GeForce GTX 1070 - 1080 1 2 2 776-1262 518-790 1 658 720 8GB 27 123% 22 Dual NVENC Cards gives double FPS.

4 Except P4000 quadro P4000 1 U 1 388-631 259-395 1 658 720 8GB 27 128% 21 quadro P5000 1 U 2 776-1262 518-790 1 658 720 16GB 55 314% 22 GeForce GTX 1080 Ti GP102 1 2 2 776-1262 518-790 1 658 720 11GB 38 205% 22 GeForce GTX Titan X Titan Xp 1 2 2 776-1262 518-790 1 658 720 12GB 49 223% 22 quadro P6000 1 U 2 776-1262 518-790 1 658 720 24GB 82 447% 22 Titan V Volta GV100 1 2 3 1164-1893 595-908 1 658 720 12GB 41 187% 22 I 3x NVENC , Limited by 1x NVDEC Titan RTX Turing TU102 1 2 1 446-725 595-908 1 1316 1440 24GB 82 552% 14 J HEVC 8K encoding at 30 FPS, HEVC B-Frames support and up to 25% bitrate savings for HEVC and up to 15% bitrate savings for Estimated FPS from 3rd party info GeForce RTX 2080 Ti 1 2 1 446-725 595-908 1 1316 1440 11GB 38 276% 14 GeForce RTX 2080 TU104 1 2 1 446-725 595-908 1 1316 1440 8GB 27 182% 14 GeForce RTX 2060 / 2070 TU106 1 2 1 446-725 595-908 1 1316 1440 8GB 27 182% 14 GeForce GTX 1660 Ti / 1660 TU116 1 2 1 446-725 595-908 1 1316 1440 6GB 20 135% 14 GeForce GTX 1650 Volta TU117 1 U 1 388-631 595-908 1 658 720 4GB 14 108% 13 Intel 2xxx, G440-G870 (except non GPU Models) Sandy Bridge HD 2000, HD 3000 1 U 1 120 - 1 120 - - 4 only.

5 No transcoders on Pentium or Celeron Series Intel 32xx, 33xx, 34xx, 35xx, 37xx, G16x, G2xx Ivy Bridge HD 2500, HD 4000 1 U 1 206 1 206 - 4 Poor image quality encoding Intel 41xx, 46xx, 47xx, G3xx, G18x Haswell HD 4200-4600, HD 5000 1 U 1 166 1 166 - 4 Performance regression. AVC, VC-1 and Part 2 Intel 57xx, 56xx, 68xxK, 69xxK+X (except non GPU Models) Broadwell HD 5300-6300P 1 U 1 1 - 4 VP8 hardware decoding Intel 6xxx Only Skylake HD510, 530, 580 1 U 1 150 1 150 - 4 8bit Support Intel 7xxx-8xxx (except 79xx) (not 7640X, E3-1xx0) Kaby, Coffee, Wiskey Lake HD610-HD640 1 U 1 1 - 4 Main10/10-bit VP9 8-bit and 10-bit Intel (except non GPU Models) Ice Lake Gen11 1 U 1 1 - 4 VP9 8/10-bit HDR10 Tone Mapping Chips H264 HEVC Chips H264 HEVC VRAM Graphics Card Model Chipset Family Chip Model # of Chips Session Limit NVENC FPS NVDEC FPS Streams for VRAM* VRAM Streams Card Load * (Less than 100% Best) Recommended Plex Streams VDPAU * FPS Based on 1080p@20 MBPS YUV 4:2:0 8-bit.

6 HEVC ( ) FPS based on 1080p@20 MBPS 8-bit *VRAM listed are common sizes. * NVENC FPS based on Single Pass quality profile. (High Quality High Performance) *VDPAU NVIDIA PureVideo Information [ #Nvidia_VDPAU_Feature_Sets] VRAM Bandwidth Preliminary testing sees a single NVDEC job on a 128 Bit GTX 1050 Ti 4GB (Pascal) unable to use more than 30% (112 FPS) of the NVDEC . Two streams hold about 50%, Three about 80% and more than four streams to reach 100% NVDEC saturation. A 256 Bit GTX 970 4GB (Maxwell 2nd Gen) can hit 100% NVDEC saturation (376 FPS) with a single stream. The difference between the 128 Bit, 192 Bit and 256 Bit Memory bandwidth needs further testing. It looks like a 128 Bit memory bus will not cause performance issues with multiple transcodes but will see Plex offline Sync jobs only able to use 30% of the NVDEC chip. VRAM Stream Card Load Card load is calculated from the smallest of the NVENC FPS and NVDEC FPS then divided by Streams for VRAM combined FPS to provide card.

7 For example; quadro K2200 4GB model can NVDEC 1080p@376 FPS and NVENC 720p@502 FPS and fit a maximum 16 transcodes in its 4GB VRAM. NVDEC being the smaller FPS we take 16x @30 FPS equals 480 FPS, more than the NVDEC can process. This card will be under 128% Load to deliver 16 streams causing buffering but could deliver 12 streams at 96% load. See Recommended Plex Streams for more information. Exceeding VRAM Exceeding the VRAM usage will cause new transcodes to buffer indefinitely, even once VRAM has dropped below the maximum. The Plex Client will need to stop the play request and request it again once VRAM usage has dropped. Choose a card with enough VRAM to avoid this. Streams for VRAM Streams for VRAM is how many Plex Streams will fit in VRAM at any one time, this figure is based on 1080p@15 MBPS to 720p@4 MBPS per stream. See Recommended Plex Streams and Exceeding VRAM for stream buffering issues.

8 TRANSCODE PROFILE KEPLER MAXWELL 1ST GEN MAXWELL 2ND GEN PASCAL & VOLTA 720P@3 MBIT TO SD 2 MBIT 102MB 207MB 190MB 720P@3 MBPS TO 720P 2 MBPS 104MB 190MB 195MB 720P@6 MBPS TO 720P@2 MBPS 98MB 218MB 170MB 720P@6 MBPS TO 720P@4 MBPS 108MB 229MB 190MB 1080P@15 MBPS TO 720P@2 MBPS 122MB 231MB 280MB 1080P@15 MBPS TO 720P@4 MBPS 133MB 248MB 300MB 1080P@15 MBPS TO 1080P@8 MBPS 142MB 250MB 320MB 4K@68 MBPS TO 720P@4 MBPS - - 1220MB 4K@68 MBPS TO 1080P@8 MBPS - - 1290MB Recommended Plex Streams Maximum 1080p@15 Mbps streams to maintain live streams without buffering pauses based on FPS of chipset and VRAM available. This calculation is based on 1080p@15 MBPS to 720p@4 MBPS per stream. See Streams for VRAM for more precise usage. Multiple NVENC Chips Some cards have either multiple NVENC chips or multiple Graphics chips each containing an NVENC chip. Taking into account that you need to decode the same FPS as you encode let s look at the GTX 1070; It supports NVENC @1262 FPS but only NVDEC @658 FPS, this would mean 658 FPS@1080p is the maximum throughput with a single pass encoding profile, not the full 1262 FPS.

9 Using a high quality profile the GTX 1070 supports NVENC @776 FPS, you will see improved image quality and smaller encoded streams while still reaching the 658 FPS@1080p throughput. Input streams lower than 1080p or that do not use NVDEC (eg. Unsupported CODEC) can allow this card to NVENC over the NVDEC limit. HEVC NVDEC on this card supports 720 FPS meaning a HEVC to workload would yield a transcode of 720 FPS, 60 FPS better than to GPU Power Usage Without an attached monitor, a GTX 970 with a max draw of 145w will draw about 80w with 100% NVDEC to NVENC transcoding. Microsoft Windows Session Limits can be overridden in Windows [ ]. This can be a tricky setup, a more sturdy option choose a quadro Card with no session limits. Linux Session Limits can be overridden on GeForce and quadro Cards [ ] Please make sure the card has enough VRAM for this. NVDEC for Linux needs or newer and a patch.

10 [ ] Please make sure the card has enough VRAM for this. Linux NVENC / NVDEC Monitoring Live usage stats: watch -n 2 ' NVIDIA -smi -q -d UTILIZATION' Usage Report: NVIDIA -smi Plex Media Server Optimisation Run Plex Library from an SSD (NVMe SSD for best performance) by making a symlink in Linux sudo ln -s "/SSD/Plex Media Server" "/var/lib/plexmediaserver/Library/Applic ation Support/Plex Media Server" Move the Transcoder temporary directory to a RAM disk using the Linux default: /dev/shm /dev/shm allows by default 50% of RAM to be used as cache You will need about 1GB per 1080p stream for a 900 sec Transcoder default throttle buffer VDPAU Reference VP4 C Mar-11 Supports complete acceleration for MPEG-1, MPEG-2, MPEG-4 Part 2 ( MPEG-4 ASP), VC-1/WMV9 and Global motion compensation and Data Partitioning are not supported for MPEG-4 Part 2. VP5 D Apr-11 Similar to feature set C but added support for decoding with a resolution of up to 4032 4080 and MPEG-1/MPEG-2 with a resolution of up to 4032 4048 pixels.